Improve NEMOv4 numerical reproducibility

[sergisiso]

NEMOv4 is still not tested for full numerical reproducibility as we do for NEMOv5. We just grep and compare the most significant digits of run.stat

Currently NEMOv4 is compiled with aggressive flags hardcoded in the runner NEMO/arch/* folder, this include -O3 -Mflushz -Mfma -Minline -fastsse, it would be good to parameterise this so that they can be explicitly controlled in the workflow and use the reproducible -O2 -Mnovect -Mnofma -gpu=math_uniform.

Also a few files of NEMOv4 did not work with "teams loop" but do with "loop", I suspect it has different behaviour with characters, but we need better understanding of this (do the flags above matter?).